+ Reply to Thread
Results 1 to 6 of 6

Thread: Ordinal or Multinomial Regression

  1. #1
    Points: 69, Level: 1
    Level completed: 38%, Points required for next Level: 31

    Posts
    6
    Thanks
    2
    Thanked 0 Times in 0 Posts

    Ordinal or Multinomial Regression




    Hi,

    If I have a variable called time spend following politics with the following possible responses:

    No time
    under half an hour a day
    between half an hour and one hour
    between one and one and a half hours
    between one and a half and two hours
    over two hours

    I want to look at predictors of this model (all nominal variables). I am not sure whether a ordinal or multinomial regression works best for this. Could someone please advise.

    Thanks

  2. #2
    TS Contributor
    Points: 12,227, Level: 72
    Level completed: 45%, Points required for next Level: 223
    rogojel's Avatar
    Location
    I work in Europe, live in Hungary
    Posts
    1,470
    Thanks
    160
    Thanked 332 Times in 312 Posts

    Re: Ordinal or Multinomial Regression

    Hi,
    the Likert scale seems to be an approximation of a continuous variable - so AFAIK ordinal regression would be better though it takes more samples.

    http://stats.stackexchange.com/quest...gression-model

    regards

  3. #3
    Fortran must die
    Points: 58,790, Level: 100
    Level completed: 0%, Points required for next Level: 0
    noetsi's Avatar
    Posts
    6,532
    Thanks
    692
    Thanked 915 Times in 874 Posts

    Re: Ordinal or Multinomial Regression

    There is a formal test for the legitimacy of using ordered logistic regression - SAS calls this the proportional odds assumption. If you reject the null then you use multinomial rather than ordered regression. However, with large samples and many IV you can reject the null when you should not. Allison suggest you need at least ten cases for each level of the DV [here 60] but I am guessing you have that many cases
    "Very few theories have been abandoned because they were found to be invalid on the basis of empirical evidence...." Spanos, 1995

  4. The Following 2 Users Say Thank You to noetsi For This Useful Post:

    DGM1 (03-28-2016), rogojel (03-28-2016)

  5. #4
    TS Contributor
    Points: 12,227, Level: 72
    Level completed: 45%, Points required for next Level: 223
    rogojel's Avatar
    Location
    I work in Europe, live in Hungary
    Posts
    1,470
    Thanks
    160
    Thanked 332 Times in 312 Posts

    Re: Ordinal or Multinomial Regression

    hi noetsi,
    maybe it would make sense to run the model and see how well it fits, first? If it does not, then probably the proportional odds assumption is not verified.

    regards

  6. #5
    Fortran must die
    Points: 58,790, Level: 100
    Level completed: 0%, Points required for next Level: 0
    noetsi's Avatar
    Posts
    6,532
    Thanks
    692
    Thanked 915 Times in 874 Posts

    Re: Ordinal or Multinomial Regression

    I don't think fit has anything to do with the proportional assumption. Its an assumption that you are testing, I forget which one, not how well the model fits the data. In theory, if commonly not in fact, you test model assumptions then look at the results.

    Of course in reality if the results are no good who would waste time on the test of assumptions
    "Very few theories have been abandoned because they were found to be invalid on the basis of empirical evidence...." Spanos, 1995

  7. #6
    Points: 69, Level: 1
    Level completed: 38%, Points required for next Level: 31

    Posts
    6
    Thanks
    2
    Thanked 0 Times in 0 Posts

    Re: Ordinal or Multinomial Regression


    Thanks for the helpful reply. Yes I have heard about proportional odds and how stringent the test is. My understanding was that you could make a dummy for each of the possible outcomes and run normal logit models, then if the odds ratios for the predictors were different across the models (6 in my case) then you could be more sure that the proportional odds assumption has actually been violated. Any thoughts on this?

    Also, if possible, would it be possible to tell me how my interpretation of the model would change. For example, if I am comparing high political interest to a low political interest (IV) against each of the different time periods for following politics of TV (comparison group = watching no TV), is it possible to say that if the odds ratio for watching 31-60 mins is larger than watching 1-30 mins in comparison to watching no TV then there is a greater likelihood of watching 31-60 mins? Or am I just limited to saying that if the odds ratios are all over 1 then there is an increased likelihood for all lengths of TV usage in comparison to no TV.
    Last edited by DGM1; 03-28-2016 at 06:56 PM.

+ Reply to Thread

           




Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts






Advertise on Talk Stats